Lead research and scoring that your sales team will actually trust

What this actually solves

Most of a sales development day is not selling. It is looking things up. Opening the website, finding the headcount, working out whether they are the size you serve, checking whether they already run something that competes with you, reading enough to write an opening line that does not sound like a template. That research is genuinely necessary — a message sent without it gets deleted — and it consumes the majority of the hours of the most expensive activity in the funnel.

Then most of that research is discarded, because the majority of leads do not qualify. The work of finding that out is identical to the work of qualifying, so a large share of the day is spent producing the answer no. Meanwhile the leads that would have said yes sit in the queue, and response time is one of the few things in this process that reliably predicts whether a conversation happens at all. A form filled in on Friday evening that gets a reply on Monday afternoon has already cooled.

What a qualification system changes is the order of operations. The research happens immediately and uniformly on every lead, the disqualifications are made before anyone spends attention on them, and what reaches a person is a shortlist with the reasoning attached and a draft to work from. The team still decides and still sends. They just stop spending their day producing the answer no.

How the work runs

1. Define the profile, including what disqualifies. Industry, size, geography, technology in use, funding, hiring signals — and just as explicitly the companies that are never a fit, which is usually the more useful half of the conversation and the half that has to be worked out in the room rather than read off an existing document. Three to five days.
2. Connect the sources. Company website, professional profiles, an enrichment provider such as Apollo or Clay, and your own CRM so existing customers and open opportunities are recognised rather than researched again from scratch. Three to five days.
3. Build the research agent. A defined set of tools rather than open-ended browsing, each returning a specific piece of information, with the output constrained to a JSON structure — approximate size, apparent current tooling, hiring signals, and two or three concrete observations usable in a first message. Constrained output is what makes the result checkable. One to two weeks.
4. Build scoring in two layers. Hard rules in code, applied first, rejecting outright. Soft signals evaluated by the model with agreed weights, producing a score across three bands — hot, warm, reject — and, critically, a written reason for every one. Four to six days.
5. Write the message generation. Tight constraints — under ninety words, one real specific about their business, one clear question, no flattery — with drafts landing in the CRM awaiting approval rather than going anywhere. Four to six days.
6. Write back into the CRM and build the review loop. Score, band, reasoning and research summary attached to the record, plus the weekly view where the team marks the scores that were wrong so the weights can be adjusted. Three to five days.
7. Calibrate against history before it touches a live lead. Run the scoring over past leads whose outcomes you know and check that the ones that converted rise to the top. Three to five days.

What moves the price

The range is $1,500–$15,000 + $200/mo over 3–8 weeks, and three things move the build fee inside it.

How many sources and how deep the research goes. Reading a website and one enrichment record is straightforward. Cross-checking several sources, inferring current tooling from site signals, tracking hiring and funding activity, and reconciling contradictions between them is a substantially bigger research agent with more failure modes to handle.

CRM integration difficulty. Some systems have clean, well-documented interfaces and custom fields that behave. Others need workarounds, have rate limits that shape the design, or carry years of accumulated field customisation that has to be understood before anything can be written into them safely.

Personalisation depth and volume. One message template with a research-driven opening line is at the lower end. Different sequences per segment, per source and per band, with tone varied by role, is a larger writing and testing job — and higher lead volumes push enrichment usage and pipeline reliability requirements up with them.

What usually goes wrong

It gets wired straight to sending. The build works, the drafts read well, and someone connects the output directly to an email provider because the approval step feels like friction. Two weeks later the domain is landing in spam folders. The cause is always the same combination — volume that jumped suddenly, message bodies similar enough to be fingerprinted, and nobody watching bounces on addresses that were never validated. The system drafts; a person sends. That is a design decision, not a limitation.

A hard criterion is handed to the model. Asked whether a company is UK-based with more than fifty employees, a model will get it right most of the time and confidently wrong some of the time, usually because the website presents an international impression. Once a handful of obviously unqualified companies reach the hot band, the sales team stops believing the whole list. Absolute criteria go in code as filters; only genuine judgement calls go to the model.

Scores arrive with no reasoning. A number between zero and a hundred, with nothing explaining it, gets ignored by experienced salespeople within about a week — they cannot sanity-check it, so they fall back on their own instincts and their own list, and the system is judged a failure even though the scoring was fine. Every score carries the evidence behind it, which also makes the weekly review possible, because the team can point at the specific reason that was wrong rather than saying the scores feel off.